Wireless Communications and Networking with Unmanned Aerial
Vehicles
ISWCS 2018 - Tutorial
Walid SaadElectrical and Computer Engineering Department,
Network Science, Wireless, and Security (NetSciWiS) GroupVirginia Tech
Email: [email protected]: http://www.netsciwis.com
Personal: http://resume.walid-saad.com
Outline
Introduction and motivation Part I: Channel modeling for UAVs Part II: Performance analysis and tradeoffs Part III: Optimal deployment Part IV: Resource management for UAVs Part V: Security Concluding remarks
2
The inevitable rise of the UAV
3
Few facts: The number of UAVs will skyrocket
from few hundreds in 2015 to 230,000 in 2035
Different types of aerial objects/systems, LOS, BLOS
Includes drones, LAP, HAP, balloons, quadcopters, etc
Facebook Project AquilaGoogle Project LOON
OneWeb LEO constellation: 648 low-weight, low orbit and low latencysatellites positioned around 750 milesabove Earth …+ SpaceX from E. Musk
Matternet
Can be a small plane, balloon or drone High altitude platform (HAP) above 15 km, or Low altitude platform
(LAP) between 200 m to 6 km Proposals from Facebook, Google, spaceX to connect the unconnected
Frequency bands for HAPs: 38-39.5GHz (global), 21.4-22 GHzand 24.25-27.5GHz (region-specific)
Remotely controlled or pre-programmed flight path Control and non-payload communication (CNPC) systems
4
Unmanned Aerial Vehicles
Applications Communications, disaster management, search and rescue,
security, control, agriculture, IoT, etc Covering hotspots+ 1000x more
Advantages Adjustable altitude Potential Mobility Low infrastructure low cost Limited available energy for Drones
Also, many challenges5
Countless Applications
disaster
Coverage/capacity
V2V
Smartermobility
VR
Agriculture
6
Challenges
Deployment
Path planning/mobility
Energy efficiency
Channel modeling
Interference
Handover and moving cells
Security and privacyResource
management
Wireless Back-/Fronthauling UAV-to-UAV communication required for coordination, interference
mitigation, relaying, routing in the air, etc. Satellite and WiFi considered as candidate technologies for providing
wireless backhauling depending on latency-bandwidth requirements Satellite backhauling brings the advantage of unlimited coverage
offering the possibility of connecting the aerial network for anydistance However, the latency introduced by the satellite links (GEO) may affect
some real time services such as voice and real-time video. To avoid satellite delays and the cost, WiFi links can be used albeit
reduced coverage and capacity (doubtful QoS guarantees..)Recent interest in Free Space Optics
License free PtP narrow beams But tackle rain, fog and cloud attenuations Multi-connectivity to the rescue..? 7
Backhaul
8
Tools Usefuls for UAVs
5G+5G+
Physics• Mean field
• Random graphEconomics• Matching theory
• Pricing
Game theory (GT) and learning • Decision making
• Resource management• Clustering
• Supervised, non-supervised learning
Control Theory• Lyapunov• Consensus
Stochastic geometry• BS/UE location
Stochastic optimization• CSI/QSI uncertainties
Random matrix theory• Asymptotics
Transport Theory• Association
• Mobility
In this tutorial, we will (briefly) touch on GT, optimal transport, and learning
9
Part I –Air-to Ground
Channel Modeling for UAVs
10
Air-to-Ground AtG Channel Model Radio propagation in AtG channel differs from terrestrial
propagation models Typically radio waves in AtG channel travel freely without obstacles
for large distances before reaching the urban layer of man-madestructures.
UAV-ground channels typically include: Line-of-sight (LOS) and NLOS links A number of multi-path components (MPC) due to reflection, scattering, and
diffraction by mountains, ground surface, foliage
Common models define a LOS probability between UAV and ground user that depends on:
Environment (suburban, urban, dense urban) Height (h) and density of the buildings (building/km2)
11
Air-to-Ground Channel Model Received signals include:
Line of sight (LOS): strong signal (G1) Non-line of sight (NLOS): strong reflection (G2) or fading (G3)
Each group with a specific probability and excessive loss Dominant components
LOS links exist with probability P and NLOS links exist with probability 1-P
Consider LOS/NLOS separately with different path loss values Excessive path loss
sampleshistogram
12
Air-to-Ground Channel Model Model by Al-Hourani et al. Buildings and environment impact the propagation
Distribution of buildings’ heights:
Suburban Urban Dense urban Highrise urban
A scale parameter depends on environment according
to a Rayleigh pdf
13
Ray Tracing Simulation Allows the prediction of signal strength in an accurate
manner Based on a simulation of actual physical wave
propagation process Can consider different ray types: Direct, Reflected and
Diffracted rays
Requires buildings database
3D predictions
14
Ray Tracing Simulation Propagation Group Occurrence Probability, obtained at
frequency = 2 GHz for an urban environment Group 1: LOS Group 2: NLOS
Example of a group occurrence curve fitting for two groups
Occurence probabilityof a certain propagationgroup at a certain angle
Parameter depends on environment
Back to LOS Probability In urban environments, the LOS probability is given as:
15
ratio of built-up land area to the total land area mean number of buildings per
unit area (buildings/km2 )
Antenna height…For large values of h, P(LoS) is acontinuous function of θ and environmentparameters see next slide
16
LOS Probability approximation Probability of having LOS link:
Trend approximated to a simple modified Sigmoid function (S-curve)
Increasing LOS probability by increasing elevation angle or
UAV’s altitude
B and C: constants that dependon the environmentθ: Elevation angle
17
Shadow Fading Modeling shadow fading
Received signal power
Shadow fading
Gaussian distribution
Parameters depend on environment
18
Ricean channel model Small scale fading is described by the Rician distribution
due to the presence of a strong LOS component in theAtG channel
Distribution of the received signal amplitude:
Rician K factor: Depends on the environment Lower for denser environments
LOS amplitudeAverage multipath component power
Bessel function
L-BandRicean K-factors = 12 dB and 27.4 dB inC-band in the near-urban environment.
14 dB in L-band and 28.5 dB in C-band for thesuburban settings.
Way forward Air-to-air channel models (still lacking in literature) The probabilistic model may not be the best, real-
world measurements can help Airframe shadowing for large-sized or small-sized
aircraft, tree/building shadowing at low altitude small UAV, also terrain shadowing for mountainous scenarios or beyond LOS conditions Of relevance here are the works of Matolak and NASA
How to integrate multiple antennas, what is the most adequate number of elements and their location (MIMO or mmWave air-to-ground channels?)
19
20
Part II –Performance
Analysis
21
Drone small cells in the clouds: Initial insights on design and
performance analysis
System Model
Downlink scenario Drones provide coverage for a target area Scenarios:
Single drone 2 drones without interference 2 drones with intercell interference
Target: Meeting the minimum SINRrequirement on the ground
22
Main Goals
Determining the optimal altitude of drones Leading to maximum coverage Full coverage using minimum transmit power for the drones
Optimal deployment of two interfering drones Distance between the drones? Altitudes?
Highlighting tradeoffs while deploying drones Interference, coverage, transmit power
23
Impact of Drone’s Altitude
Higher altitude: Higher path loss vs. higher LOS proba. Lower altitude: Lower path loss vs. more NLOS Altitude and flight constraints
Higher and lower altitudes are bounded
24
What is the Optimal
Altitude?
Single Drone
Minimize transmit power via an optimal altitude Path loss as a function of elevation angle:
25
Environmental parameters
Additional loss for NLoS
Optimal altitude
Optimal Altitude
Optimal altitude depends on the area size (Rc) Increasing drone’s altitude to service larger areas
26
@Low-altitude: high shadowing+ low LOS probabilitycoverage radius decreases
@ high-altitude: high LOS probability but PLIncreases –> Coverage decreases
E.g.; optimal altitude for providing 500m coverageradius while consuming min. tx power is 310 meters
Altitude increases w/ coverage radius
Two Drones
Given a desired geographical area: Maximize the total coverage area What is the distance between drones? What is their altitude?
27
Total coverageDistance between drones
No Interference Case Deploying each drone at its optimal altitude Packing the coverage areas inside the target area While keeping the distance between drones as far as
possible, but inside the target area
28
Maximum coveragerange of each drone
Total coverage
Two Interfering Drones Consider a rectangular geographical area High distance between drones: covering undesired area Small distance between drones: high interference
29• No coverage in between due to the interference
• Drones should not be placed too close
• 300 meter altitude• 1100 meter separation
Results
Bounded target geographical area Existence of optimal drones’ separation distance for
maximum coverage
30
At high drone distance, although separated,coverage ratio is low (undesired)
Likewise, if too close interference increases.
optimal separation distance exists!
31
Unmanned Aerial Vehicle with Underlaid Device-to-Device
Communications: Performance and Tradeoffs
System Model
Downlink Scenario: UAV coexists with a device-to-device (D2D) network
Two types of users: downlink users (DU) and D2D UAV provides service for downlink users Interference between UAV and D2D transmitters Static and Mobile UAV Cases
32
UAV and D2D: Assumptions
Users (DU and D2D) distributed based on Poisson point process (PPP) Number of users follows Poisson distribution, but uniformly
distributed over the area The number of points in a bounded area has a Poisson
distribution with mean e.g. λ×A or λ×B
Underlay D2D communications:use existing licensed spectrum
Can we analyze the performancetradeoffs for UAV deployment
33
Derive the average coverage probability and sum-rateexpressions Finding the relationship between UAV parameters (altitude,
etc.) and rate/coverage Finding some fundamental performance tradeoffs
What is the optimal altitude of the UAV that maximizes the coverage and rate? Fundamental tradeoffs between DU and D2D users
How to optimize coverage using UAV mobility ?
Main Objectives
34
Coverage probability for downlink users (DUs)
Coverage probability for D2D users
Average rates
Performance Evaluation Metrics
35
SINRSINR ThresholdPolar coordinates
Static UAV: Analytical Results
D2D Coverage Probability
DU Coverage Probability
36Interference from D2D links
UAV transmit
power
D2D transmit
power
UAV-D2Ddistance
DistanceBetween
D2D pairs
D2Ddensity
LoSprobability
Number of D2Ds Impacts interference generated at the DUs
Distance between each D2D pair UAV’s location and altitude
Impacts air-to-ground channel Transmit powers of D2D and UAV
Directly affect the coverage probabilities SINR threshold Overall, we have a tractable expression to analyze UAV
coverage
Key parameters
37
Results: Static UAV Optimal altitude for DU maximum coverage
LoS and NLoS tradeoff
Impact of altitude on D2D coverage probability UAV is an interference source for D2D
38
Optimal
Results: Static UAV Average sum rate vs. altitude
Considering DU and D2D rates Depends on the distance between each D2D pair 𝑑
39
The lower is 𝑑 the higheris the sum-rate
Mobile UAV UAV moves over the target area Transmits at given geographical locations: “stop points”
Goal: satisfy DUs coverage requirements by coveringthe entire area
Analyze the impact of UAV’s mobility on the outageprobability of D2D links Considering the spatial correlation in D2D communications
Question: What is the minimum number of stop points(delay)? 40
Mobile UAV Minimum number of stop points
Depends on: UAV altitude, D2D density, size of area,coverage constraint
Moving the UAV to provide complete coveragefor the area of interest Using optimal circle covering approach Full coverage with minimum
number of circles
41
Results: Mobile UAV Maximum coverage radius vs D2D density
Higher number of D2Ds: higher interference Decreasing coverage radius!
42
Results: Mobile UAV Number of stop points vs. D2D density
Higher number of D2Ds: higher interference Increasing number of stop points!
43
Results: Mobile UAV Altitude and number of stop points
: target DU coverage requirement Altitude impacts coverage range and thus number of stop
points
44
Higher coverage requiresmore stopping points
Results: Mobile UAV Coverage-delay tradeoff
Higher number of stop points: Better coverage performance for DUs Leads to a higher delay
45
Results: Mobile UAV UAV affects the D2D outage
No UAV: only other D2Ds create interference With UAV: UAV+ other D2Ds are interference sources
Moving UAVs leads to higher average outage probability forD2D network
46
47
Part III – Optimal Deployment
Where and when to deploy UAV-BSs? What metrics to optimize (long term vs. short term)? How to develop wireless-aware path planning mechanisms?
Optimal Deployment and Mobility
48
UAV Base Stations (LAPs) Terrestrial Base Stations• Deployment is three-
dimensional• Deployment is two-
dimensional (with small exceptions)
• Short-term, frequently changing deployments
• Mostly long-term,permanent deployments
• Mostly unrestricted locations • Few, select locations• Mobility dimension • Fixed and static
49
Deployment strategies of multipleUAVs for optimal wireless coverage
System Model
Downlink communications
Using directional antennas for UAVs
Interference between all UAVs
Circular target area
Meeting the minimum SINRrequirement on the ground
50
Derive the coverage probability and coverage range of each UAV
Maximize the coverage performance by efficient deployment of multiple UAVs
Adjust UAVs’ altitude based on antenna beamwidth
Avoid overlapping coverage to avoid interference
Main Objectives
51
Considering shadowing effect in LoS and NLoS links,
Downlink Coverage Probability
52
Received signal power
Path loss
3 dB antenna gainQ function
Coverage range of each UAV:
M identical UAVs Total coverage is maximized No overlap between UAVs’ coverage areas
Multiple-UAVs deployment
53
Big circle: area of interest which needs to be covered Each small circle: Coverage region of each UAV Maximizing the packing density is equivalent to maximizing total coverage
Approach: Circle Packing Problem
54
The optimal packing of 10
circles in a circle
The optimal packing of 15 circles in a
square
The optimal packing of 6 circles in a right isosceles
triangle
Coverage radius vs. number of UAVs (circle packing):
Upper bound on the coverage radius:
Results
55
Altitude versus number of UAVs More UAVs:
Less coverage radius per UAV is required Reduce UAVs’ altitudes to avoid interference (overlapping)
Results
56
Meeting a total coverage requirement What is the minimum number of UAVs? Depends on the size of the area Choosing appropriate number of UAVs based on coverage
requirement and size of target area
Results
57
Total coverage and coverage lifetime tradeoff Increasing number of UAVs:
Transmit power per UAV can be reduced Higher coverage lifetime
Results
58
59
Cooperative deployment and mobility of UAVs for optimizing rate-delay
tradeoffs
Cooperative UAV Deployment
Task 3Task 2 Task 1
Task 4
Given a number of tasks in an area and some autonomous agents (e.g., UAVs) How to dispatch the agents to service the tasks? Can the agents make their own decisions on servicing the tasks? Almost no work considered the problem in the context of a
wireless/communication network Tasks are queues of data with no direct connectivity
Task 6Task 5
60
61
The problem is well studied but… Most approaches are
Robotics-oriented Mainly in military applications (tasks are targets) Other related problems (the repairman problem, dynamic vehicle
problem… ) Software engineering (autonomous agents) The tasks are usually considered as passive entities
Almost no work considered the problem in the context of a wireless/communication network With next generation self-organizing networks this problem becomes quite
relevant Nature of wireless networks (channel, traffic, etc) Quality of service
Cooperative UAV Deployment
62
Agents in Wireless Networks Given a number of tasks in an area
Consider each task as a M/D/1 queuing system generating packets with a Poisson arrival
Each task i has an arrival rate λi
The network operator, requires.. Data collection from the tasks Wireless transmission of the data to a central receiver
The network owns a number of autonomous agents that need to Decide on which tasks to service Collect the data and transmit it taking into account
The amount of data collected The delay
Task 4
Task 3
Task 5
Task 2
UAV Agents in Wireless Networks
Central Receiver(Command Center)
Task 1
How will such groups form? A cooperative game between Tasks and Agents
Solution using notions from operations research, wireless networks, and queuing theory
Collect data from task 5 and transmit it
Collect data from task 3 and transmit it
Collect data from task 4 and transmit it
Collect Data from Task 1 and transmit it
Collect Data from Task 2 and transmit it
63
64
Problem Formulation Coalitional game where
The players are the tasks and UAVs, hence, the player set N is the set of all tasks and UAVs Denote M the set of UAVs and T the set of tasks, N = M U T
Each coalition S consists of a number UAVs servicing a number of tasks t
A UAV can be either A collector: more collectors means smaller service time, less
delay Each collector i has a link transmission capacity µi For a number of collectors G servicing a task i in a coalition S the
total link transmission capacity is
A relay: more relay means better effective throughput (less outage probability)
65
Problem Statement Each coalition S can be seen as a polling system
with exhaustive strategy and switchover times Polling systems are ubiquitous in computer systems The main idea is that a server is servicing multiple
queues (sequentially or not) Exhaustive implies the server collects all the available data
from a queue before moving to the next Switchover times are the time to move from one task to
the other In this context, each coalition S consists of
A number of collectors acting as the polling system server
The tasks are the queues of the polling system Switchover time is the travel time from one task to the
next
66
Performance metrics - Delay For a polling system, it is difficult to have an exact
expression for delays, but, we can use the pseudo-conservation law for a coalition S
Stability of coalition S (polling system) requires ρS < 1 The total switchover time θS depends on the sequence in
which the tasks are visited Nearest neighbor solution to the travelling salesman problem
Utilization factor ρi: ratio of arrival rate to link transmission capacity of
collectors for a task i
Total switchover time θS of coalition S
Sum of utilization factors over all tasks in S
67
Performance metrics - Throughput For each coalition, the total effective throughput
from the data collected and transmitted is given by
Pri,BS is the outage probability for wireless transmission from task i to the central BS Improved by having UAVs working as relays on the link
between the collectors on task i and the BS
For each coalition, the UAVs and tasks are given a reward from the network operator depending on the throughput-delay trade off achieved
68
Utility function Given the throughput and delay previously defined, for
each coalition S we propose the following utility
β is a tradeoff parameter that represents the weight that a coalition puts on the throughtput and delay
The utility is based on the concept of power which is a ratio between effective throughput and delay
Utility is transferable: the total revenue achieved by coalition S with δ the revenue per unit power
Given the players set N and the utility v the question is We use the framework of hedonic coalition formation games to
solve the problem
69
Hedonic Coalition Formation In our game we can say that A UAV prefers a coalition S1 over a coalition S2 if
The UAV is not the only UAV in its current coalition S2 and The payoff he receives in S1 is higher than S2, and he had not
visited this coalition before (history tracking).
xiS is the payoff received by player i from the division of the utility
(we consider equal division for this work) h(i) history set
70
Hedonic Coalition Formation A task prefers a coalition S1 over a coalition S2 if
The payoff he receives in S1 is higher than S2, and he had not visited this coalition before (history tracking).
By using these preferences we can derive an algorithm form coalitions between the UAVs and the tasks
Having defined the preferences, the next question is How to form the coalitions?
71
Coalition Formation Algorithm Coalitions form and break as a result of
selfish decisions by the players (agents and tasks)
Switch rule
Every player switches its current coalition to join another, if and only if the new coalition is strictly preferred using the defined preferences.
72
Coalition formation algorithmInitial Network State:
Non-cooperative network
Task discovery: The central BS informs the UAVs
of the tasks in the area
Each player ( UAV or task) surveysnearby coalitions for possible switch
Each player takes the switch Decision that maximizes its payoff
Sequential switch operations until convergence
Final partition: Continuous data collection and transmission by the UAVs
The final partition is
Nash-stable, no player has
an incentive to unilaterally change its coalition
73
Simulation results (1)
74
Simulation results (2)
75
Mobile UAVs for Energy-Efficient Internet of Things Communications
System Model
Uplink IoT communications
Meeting SINR requirements of IoT devices
Periodic versus Probabilistic IoT activation models
UAVs update their locations based on devices activation patterns
76
IoT devices
Battery limited Typically unable to transmit over a long distance due to their
energy constraints UAVs can dynamically move towards IoT devices, collect the
IoT data moving IoT aggregators
Many IoT devicesinterference issue
IoT activations: Periodic: weather monitoring and smart grids applications Probabilistic: health monitoring and smart traffic control
applications.77
How to enable reliable and energy-efficient uplinkcommunications in a large-scale IoT using UAVs?
What are the joint optimal 3D UAVs’ locations, device-UAV associations and uplink power control?
Need for a framework for updating UAVs locations intime-varying networks:
1) Update times: shows how frequently UAVs update theirlocations
2) UAV trajectories
Main Objectives
78
Joint UAVs’ locations, associations, and poweroptimization
Problem Formulation
79
IoT transmit power
UAV j location
Set of active IoT devices
SINR Constraint
Channel gain
Association matrix
Decompose the problem into two subproblems Solve the problem for fixed association Solve the problem for fixed UAVs’ locations
Consider interference and non-interference scenarios separately
General Approach
80
UAVs’ locations and device-UAV association An example, given the locations of active IoT devices
Results
81 5 UAVs serving 100 active IoT devices uniformly distributed over the area
Reliability Probability that active devices are successfully served by
UAVs Significant enhancement by dynamically moving UAVs
Results
82
Total transmit power vs. number of UAVs Compared with stationary aerial base stations
Results
83
• 5% power reduction vs. baseline on the
average
Total transmit power vs. number of orthogonal channels for meeting SINR requirements More channels:
less interference and hence, lower transmit power needed to meet SINR requirements of each device
Results
84
• 100 devices served by 5 UAVs
• By increasing the number of channelsfrom 25 to 50, the total transmit powerof devices can be reduced by 68%
Time varying IoT network UAVs dynamically update their locations based on IoT
activations Probabilistic activation during [0,T]:
Beta distribution with parameters
Periodic activation: Each device has a specific activation period
IoT activation models
85
Time instance at which the UAVs’ locations and associations are updated
Depends on the activation process of IoT devices
Number of IoT devices For higher number of devices more updates are needed!
Energy of the UAVs More updates requires more mobility
UAVs’ update times
86
For probabilistic activation case Choosing appropriate update times based on
number of active devices
UAVs’ update times
87
Regularized incomplete beta function
Average number of active devices
Total number of devices
Number of devices which must be served vs. update time More frequent updates:
More devices can be served Less active (unserved) devices remain
Results
88
For a higher number of update times orequivalently shorter time period betweenconsecutive updates, the average numberof devices that need to transmit their datadecreases
Update times for different number of active devices Depends on the activation process (beta distribution parameters) Ensuring that the average number of active devices is less than a
Results
89
Average number of active devices
• To achieve lower value of a, updates need to be done more frequently so as the time interval between updates decreases.
• For e.g., to meet a = 100, 75, and 50, the 5th update must occur at t = 0.41, 0.55, and 1
UAVs update their locations according to the activity of the IoT devices
How to optimally move UAVs between the initial and the new sets of locations? Mobility with minimum total energy consumption Energy consumption of each UAV depends on travel distance,
UAV’s speed and power consumption as function of speed
UAVs’ mobility
90
Travel time
Which UAV goes where?
UAVs’ mobility
91Update time t1Update time t2
New set of UAVs’ locations
Initial set of UAVs’ locations
Energy constraint of each UAV
Can be transformed
into an assignment
problem
Transportation matrix
Energy from location k to l
Update times impact the UAVs’ energy consumption for mobility More updatesUAVs need to spend more energy on mobility
Results
92
• by increasing the number of updates from 3 to 6, the energy consumption ofUAVs increases by factor of 1.9
93
Part IV – Resource Management
Let’s first take a look on the impact of hover time
Resource management
94
UAV Networks Terrestrial Networks• Spectrum is scarce • Spectrum is scarce• Inherent ability for LoS
communication can facilitatehigh-frequency (mmW)
• Difficulty to maintain LoS poses challenges at high frequencies
• Elaborate and stringent energy constraints and models
• Well-defined energy constraints and models
• Varying cell association • Static association• Hover and flight time
constraints• No timing constraints, BS
always there
95
Optimal Transport Theory for Hover Time Optimization
Flight Time Constraints?
UAVs have limited on-board batteries Cannot fly for a long time
Flight regulations and weather conditions No-fly time and no-fly zones Wind and rain effects
Mobility based on demands Cannot stay at one location for a long time
Flight time constraints must be taken into account: Minimizing flight time while meeting the demands Optimizing the service performance under flight time constraints
96
System Model
M stationary UAVs serve N users Users’ distribution:
2D spatial distribution of users Determines how likely a user is present
M partitions each serviced by one UAV Hover time: Time duration that a UAV spends over a given area Channel model adopted is the one explained earlire Goal: finding optimal cell partitions and associations
Based on users’ distribution, hover times, and UAVs’ locations
Two scenarios: Maximizing total service data given the maximum hover times (Scenario 1) Minimizing average hover time while meeting load requirements (Scenario 2)
97
Problem Formulation (Scenario 1)
Total bandwidth for UAV i : Hover time of UAV i : Effective data transmission time: Control time which is not used for transmission:
Portion of hover time which is not used for data transmission Used for processing, computations, and control signaling. Is a function of the average number of users
Data transmitted to a user located at (x,y) served by UAV i :
98
Scenario 1 Time and bandwidth are the resources We consider some level of fairness in resource allocation:
Maximizing average total data service by optimal partitioning:
99
Depends on hover times andbandwidths
Approach: Optimal Transport Theory
Moving items from a source to destination with minimum cost
What is the best way to move piles of sand to fill up given holes ofthe same total volume?
Goal: Minimizing total transportation costs Where should each pile be moved? Our problem: transportation from users to UAVs!
100
Monge-Kantorovich Transport Problem
Given two probability distributions
Same amount of mass in source and destination What is the optimal mapping between ?
101
Back to our problem
We have a semi-discrete optimal transport problem Mapping from users’ distribution (continuous) to UAVs (discrete)
Optimal cell partitions are related to optimal transport maps
102
Finding Optimal Partitions and Associations
Finding optimal values of leads to the optimal transport mapand optimal cell partitions!
Complete characterization of partitions is now possible103
Kantorovich Duality Theorem:
Theorem 1:
Cost function depending on data service
Finding Optimal Partitions and Associations
1) F is a concave function of
2) Using gradient based method to find optimal
3) Optimal cell partitions are given by:
104
Theorem 2:
Special case: results in a weighted Voronoi diagram
Results: Scenario 1
We consider truncated Gaussian distribution for users Suitable for modeling hot spots in which users are congested
105
Results: Scenario 1
Lower : users’ distribution is more non-uniform Jain’s fairness index is one when all users receive equal service
106
Average number of users in each partition
Fairness index for average data service
Scenario 2
UAV-based communications under load constraints Goal: minimizing the average hover
time needed for serving the users By finding optimal cell partitions
107
Average hover time of UAV i to service partition :
Transmission time Control time
: rateLoad (in bits)
Problem Formulation (Scenario 2)
Average total hover time of UAVs:
We will characterize the optimal solution using optimaltransport theory again
108
Optimal Partitions
Proof idea: Proving the existence of solution Comparing optimal partitions and a non-optimal variation of those Then characterizing the solution
Note: weighted Voronoi is a special case (with no control time)109
Theorem 3: optimal cell partitions can be characterized as
Results: Scenario 2
Average hover time vs. control time
110
Results: Scenario 2
Hover time and bandwidth tradeoff
111
112
Beyond 5G with UAVs: Foundations of a 3D Wireless Cellular Network
System Model 3D aerial network: Drone-users (drone-UEs) Drone base stations (drone-BSs) HAP drones for wireless backhaul
Important metrics: Connectivity Latency
Two key problems: 3D network planning of drone-BSs
Deployment and frequency planning 3D cell association for drone-UEs 113
Proposed Framework
114
3D deployment of drone-BSs and frequency planning:
truncated octahedron cells
Estimating the spatial distribution of drone-UEs using
machine learning tools
Optimal 3D cell association forminimum latency of drone-UEs using optimal transport theory
Drone-BSs’ locationsand co-channel cells
3D spatial distributionof drone-UEs
Network Planning of Drone-BSs Inspired by 2D hexagonal cells Hexagons covers an area without gap or overlap Closest to circle
Omni-directional antenna
How about in 3D?Criteria:
Full coverage with minimum number of drones Closest shape to a sphere Tractable Candidates for regular 3D shapes:
Cube, Hexagonal prism, Rhombic dodecahedron, Truncated octahedron
115
Results: Network Planning
116
Number of drone-BSs needed for full coverage of space Different space filling polyhedra
3D Network Planning of Drone-BSs
117
Truncated octahedron structure will provide an initial way to place drone-BSs Placing drone-BSs at centers of truncated octahedrons
14 faces:8 hexagons6 squares
Deployment and Frequency Planning
118
Theorem 1. the three-dimensional locations of drone-BSs are:
Theorem 2. the feasible integer frequency reuse factors can be determined by:
where a, b, c are integers chosen from set {…,-2,-1, 0, 1, 2,…}
n1, n2, n3, m1, m2, and m3 are integers thatsatisfy above equations
Integer frequency reuse factors: 1, 8, 27,64,…
Results: Frequency Planning
119
Integer frequency reuse factors (q): 1 and 8
Higher q : higher SINR but requires more bandwidth
Latency-Minimal 3D Cell Association
120
Latency in serving drone-UEs Transmission latency Backhaul latency Computational latency
Depend on: resources, congestion, and 3D cell association
Transmission Backhaul ComputationAverage number of independent drone-
UEs in cell n
3D cell partition
Drone-UEs’ distributio
n
Packet length
Bandwidth
Total number of drone-UEs (assumed to be large) Challenging to solve
Solution Characterization
121
Using tools from optimal transport theory Finds optimal mapping between two probability measures Considering a semi-discrete optimal transport problem
Mapping drone-UEs’ distribution (continuous) to drone-BSs (discrete) Optimal 3D cell partitions are related to optimal transport maps
??
Steps: - Existence of solution by the existence an optimal map- Comparing optimal partitions and a non-optimal variation of those- Characterizing the solution
Solution Characterization
122
Theorem 3: the optimal 3D cell partitions are characterized by:
Note: 3D cell shapes depend on: - drone-UEs’ distribution, drone-BSs’ locations, backhaul rate, computational speed
Results: 3D Cell Association
123
Proposed approach vs. SINR-based association Reduces latency Improves spectral efficiency
Results: Latency
124
Latency increases by increasing packet size Transmission Computation Backhaul
M. Mozaffari, A. Taleb Zadeh Kasgari, Walid Saad, Mehdi Bennis, Merouane Debbah, “Beyond 5G with UAVs: Foundations of a 3D Wireless Cellular Network”, https://arxiv.org/abs/1805.06532
125
Caching in the Sky: Proactive Deployment of Cache-Enabled Unmanned Aerial Vehicles for Optimized Quality-of-
Experience
System Model Considerations:
Users mobility Users’ content request Caching at UAVs UAVs’ deployment
Transmission links (a) Content server->BBUs->RRHs->users (b) Content server->BBUs->UAV>users (c) Cache->UAVs->users 126
Maximizing users’ quality of experience (QoE) using minimum UAVs’ transmit power
Optimizing
Users association
UAVs’ locations
Content caching
Main Objectives
127
General Approach
For learning and predictions, we use the neural network
framework of echo state networks 128
Notion of “reservoir” (random) Only need to train the output layer
via linear regression Good at dealing with time stamped data
Echo State Networks
W in
Wx n
y n
W
y n1 y n
W in W
Wout Wout Wout Woutx n1 x n
WW in W in W
y n2 x n2
u n u n u n1 u n2
Echo State Networks
Step 3.
Training Process
Step 1.
Step 2.
Usage Process
131
ESN for Caching ESN model consists of
Agents: Baseband units of a CRAN Input: the input is the users locations and context
information (e.g., requested videos, etc.) Output: the output is prediction of mobility patterns ESN model: This is the reservoir model, without going
through it now, it is composed of a set of matrices that enable the recurrent neural network learning/predictions
Conceptor: use of a week mobility as “pattern”
For simulations, we use real data from BUPT and the Youku video website
Average transmit power of each UAV vs. number of users Using proposed approach, 20% reduction in transmit power
compared to other algorithms
Results
132
The percentage users with satisfied QoE versus the number of the users
Using UAVs leads to a significant QoE improvement!
Results
133
Decreasing transmit powers while increasing the number of storage units UAV will directly transmit the requested contents to the users
Results
134
135
Liquid State Machine Learning for Resource Allocation in LTE-U UAVs
System Model
Consider the downlink of an LTE-U network composed of K dual-mode UAV-base stations and W ground WiFi access points
The UAVs are equipped with cache storage units UAVs can be deployed as flying base stations with caching capabilities The UAV can cache a set of C popular content that can be pre-fetched from
a local cloud Cloud-UAV fronthaul links are licensed, wireless links
On the licensed band, we consider an FDD mode for the downlink ofthe LTE-U users, while we use a TDD mode with duty cycle for theunlicensed band LTE-U transmissions will happen for a fraction of time over the unlicensed
band, and will be muted for the rest of the slot
The ground WiFi access points (WAPs) use a standard CSMA/CA136
WiFi Data Rate Model The WiFi saturation capacity over the unlicensed band will be:
Tc, Ts, and Tσ represent the average time thechannel is sensed busy because of a successful transmission, during a collision, and the duration of an empty slot, respectively Computed using conventional approaches The WiFi network uses a standard DCF and RTS/CTS access schemes
The per user WiFi rate will be:137
# users Probability of occurrence of a transmission Successful
transmissionprobability
Averagepacket
size
UAV Data Rate Model We use the air-to-ground channel model introduced by Hourani et al.,
in which the probability of a LoS connection depends on the ground environment, and, thus, the average path loss will be:
with
and
The data rate on the licensed band will therefore be:
138
Fraction of licensedband for user i Bandwidth Fronthaul power
UAV Data Rate Model Over the unlicensed band, the data rate of the UAV will be:
The fronthaul UAV k-cloud rate for each associated user will be:
139
Number ofusers at t Average path loss
Fraction of time forunlicensed band
Queuing Model The queue length of user i at the start of slot t will be:
The data rate will be
Link (a) is the UAV-user link over the licensed band Link (b) is the UAV-user link over the unlicensed band Link (c) is the cloud-UAV-user licensed band link Link (d) is the cloud-UAV-user unlicensed band link 140
Queue length Arrival rateData rate
Problem Formulation Queue stability will be used to measure the delay:
The key goal is to solve the following resource management problem:
Challenging problem because it includes both content predictions/caching and spectrum management which is non-convex and complex
Solution? Neural networks for predictions AND resource management!141
Liquid State Machines We need an algorithm that can: a) track the network over time, b)
store user information, and c) rapidly find the resource management solution We use spiking neural networks (SNNs) since they can capture accurate
activation of neurons which enhance their predictive capabilities SNNs have two major advantages: fast real-time decoding of input
signals that are continuous and a temporal dimension that can help a high volume of information for predictions
However, general SNNs are computational complex to train Solution via liquid state machines (LSMs) LSMs are SNNs that are easy to train as they use the concept of
reservoir computing (basically random training) to make them amenable to easy implementation
142
LSM for Predictions Basic architecture of LSM
The “liquid” is a leaky-integrate-and-fire (LIF) SNN that mimics exactly a biological neuron
The input in our model is which is a vector that represents the users’ context information
The output is a request distribution vector The output function builds the relation between LSM state and the
content request distribution 143
The output function is trained in an offline manner using ridge regression:
Then, the prediction of the output can be found:
We now need to define another LSM for solving the resource management optimization problem
LSM for Predictions
144
LSM state sequence
Identity matrix
Learningrate
Targetoutput
The UAVs are the agents that run the LSM for resource management The input is a vector mk(t) of actions observed by UAV k on other
UAVs, with each action being a user association scheme Using this input and one of our previous results, we can recast on
cached content, we can recast the original optimization as a convex problem to choose the actions
The output of the LSM is a vector bk(t) that provides the resource allocation results, with each element being the expected number of stable queue users:
This is used with the output function to solve our original problem
LSM for Resource Management
145
146
Simulation Results
Real data fromYouku
LSM provide veryaccurate predictions
147
Simulation Results The average number
of stable queue usersincreases with network size
Caching brings aboutsubstantial gains,even without LSM
LSM provides furthergains
148
Simulation Results
The proposed LSM algorithm leveragethe power of SNNsto substantiallyreduce convergencetime (about 1/3 lessthan Q-learning)
149
Simulation Results
The more we cancache, the more users we can serve withstable queues
150
Cellular-Connected UAVs over 5G: Deep Reinforcement Learning for Interference
Management
System Model Uplink of a cellular network composed of S base stations
(BSs), Q ground users, and J cellular-connected UAVs UAVs must co-exist with ground users and share resource
blocks UAVs are assumed to be flying at a constant altitude (different
for different UAVs) and collecting data (e.g., surveillance,sensing, etc.) that needs to be transmitted to the ground BSs Each UAV has a specific mission and needs to move from an
origin to a destination while transmitting data along the way For ease of exposition, we consider a virtual grid that the UAVs use
for their mobility, i.e., they move along the centers of small grids Areas within the grid are chosen to be sufficiently small
151
The SINR for UAV j’s transmission to a ground BS s, over RB c is:
The achievable rate for a UAV j will then be given by:
We also consider queuing latency, using an M/D/1 model:
UAV-BS Transmission Model
152
Total Interference (ground and air)
Bandwidth
RicianchannelUAV power
# RBs
Packetarrivals
Data rate
Ground Users Data Rate Model For the ground users, the achievable data rate will be given by:
Ground users can potentially be significantly affected by interference stemming from flying UAVs (due to the drones’ better channel towards the ground BSs)
Our objectives will therefore be to answer the following key questions: How can we design a wireless-aware path planning mechanism for
cellular-connected UAVs? How can the designed path plan optimize the UAVs’ mission
goals, while minimizing impact on the ground network? 153
Total Interference(ground and air)
Problem Formulation We can pose our path planning problem as follows:
154
Tradeoff between interferenceto ground, delay, and path length
Each area is visited once
Maintain origin-destination
Arrive/leave same area
One BSper UAV
SINR/powerconstraints
Problem is challenging to solve in a centralized manner, especially to do joint power allocation, navigation, and cell association
Objective functions are coupled through interference => a game-theoretic approach is appropriate!
We formulate a dynamic game:
The utility functions can be defined as follows:
Game-Theoretic Approach
155
UAVs Stages Actions
State space: distance/orientation
Distributions Utility functions
Lagrangian conversion of centralized case
Solution Approach Since the game is dynamic and has a large action space, it is
challenging to analytical characterize the subgame perfect Nash equilibrium (SPNE) Such characterization may also require full knowledge of the
system and state, which is not very practical We will seek to develop a reinforcement learning (RL) algorithm
that enables the UAVs to autonomously find the SPNE RL algorithm with predictive capabilities is needed to operate with
minimal information Actions are time varying => need dynamic RL predictions and
highly adaptive algorithm Recurrent neural networks!
156
157
ESN for UAV We not only need to deal with time-stamped data, but
also with large action sets We will propose a novel deep ESN architecture Input: the input to the first layer is the external network state
while input to subsequent layers are previous layers Output: the output is estimation of utility function ESN model: This is the reservoir model, without going
through it now, it is composed of a set of matrices that enable the RNN learning/predictions and is trained by our network state
When it converges, the algorithm will find an SPNE, but establishing general convergence is challenging
158
Simulation Results
Proposedwireless-awareapproachavoids causingground interference
159
Simulation Results
160
Simulation Results
161
Simulation Results
Convergence depends on learning rate (0.01 is ideal for this case)
Other UAV Comm. Approaches UAVs as backhaul (see U. Challita and W. Saad, GLOBECOM
2017) More on machine learning (see M. Chen, W. Saad, et al.,
GLOBECOM 2017) UAVs as relay stations (see works by L .Swindelhurst et al. and
R. Zhang et al.) Cyclical resource allocation with optimal deployment of UAVs
as relays (see works by Y. Zeng and R. Zhang) Deployment within a cloud radio access network and related
environments (see Yanikomeroglu et al.) Channel modeling, localization, tracking, public safety, and
related ideas (see works by I. Guvenc et al.)162
163
Part V – Security
164
CPS Security of UAVs UAVs are essentially cyber-physical systems
Cyber vs. Physical: the physical world follows (typically) laws of nature or control-theoretic models, which have different constraints and time scales compared to cyber features
Human-in the loop: man meets machine (UAV) CPS nature brings cyber and physical vulnerabilities As UAVs become more prevalent, they will face more
and more security challenges Autonomy is both a blessing and a curse Let’s see an example security problem
Delivery Drones
Drones will be used in the real-world for delivering goods or to deliver rescue mission items
165
Security of Delivery Drones
Delivery drones are prone to a variety of cyber-physical security threats Cyber attacks to hack the cyber/wireless system and re-
route the drone or to jam its communication Commercial drones will be in the range of civilian-owned
hunting rifles that can be used for physical attacks In such scenarios, humans will be in the loop!
Attackers will likely be humans (e.g., choose a high point to shoot the drone or jam its link in line-of-sight)
Vendors who own the drones will have stringent delivery times especially for medical delivery (framing effects!) 166
Basic System Model
A vendor sends a delivery drone from an origin O to a destination D In an ideal world, vendor always chooses shortest path
Presence of adversary Attackers can interdict the drone at several threat points
such as high buildings or hills to cause physical or cyber damage
A destroyed drone must be re-sent by the vendor, leading to increased delivery times and economic losses
The system can be modeled as a graph167
Basic System Model
The vendor is an evader wants to minimize expected delivery time by choosing an optimal path
The attacker is an interdictor who chooses a location to attack the drone and maximize the delivery time
Natural zero-sum network interdiction game 168
Game Formulation
Two-player zero-sum game in which both vendor and attacker want to randomize over their strategies Defender mixed-strategy vector Attacker mixed-strategy vector
Attack at location n will be successful with probability pn
The expected delivery time will be:
fh(.) is a distance function T depends on various parameters 169
Game Formulation
Vendor problem
Adversary problem As a zero-sum game, it can be transformed into two
linear programs that can be easily solved Game admits a Nash (saddle-point) equilibrium There may be more than one equilibrium, but they are all
interchangeable yielding the same delivery time But what about the human perceptions? 170
171
Expected Utility Theory
Conventionally, the Nash equilibrium is found under expected utility theory (EUT) considerations Presumes that players act rationally The players optimize the expected value over their mixed
strategies, i.e.,
Caveat: in practice, it has been empirically shown that when users are faced with uncertainty, they act irrationally
172
Are humans really rational?
How to capture such irrationality?
Example: In the real-world,
security problems often involve human decision makers at both sides of
the aisle (attack/defense) Human in
the loop
Source: Study between Kyoto University and game theorists at Caltech (June 2014)
173
Prospect Theory
Lottery example Risk impacts
how players weighcertain outcome
Uncertainty can lead players to deviate from the rational norms of EUT
Subjective perception on losses/gains In CPS and UAV, many human players are in the loop and
will have subjective perceptions on the various performance and network measures
Solution: Prospect theory!
174
Example The preferred choice between a pair (or more) of
uncertain alternatives is determined by: Value of the alternatives (as is customary) but also.. How those choices are stated!
Gain Scenario: You average monthly bill is now $450 a month. Under our new smart system your bill will now show a debit of $500 a month. Also, you may choose: A) 50% chance of a $100 credit if you join our new
wireless pricing system B) 100% chance of a credit of $50 that will keep your bill
the same
175
Example Loss Scenario: Your average monthly bill is now $450 a
month. Under our new smart system your bill will now show a debit of $400 a month. Also, you may choose: C) 50% chance of a bill for $100 if you join our system D) 100% chance of a bill of $50 that will keep your bill the
same A) and C) are identical, while B) and D) are identical Prospect theory found that people will always prefer B)
to A) and C) to D) A certain gain is preferred to an uncertain double gain! An uncertain loss is preferred to a certain, smaller loss!
176
Prospect Theory
Prospect theory Introduced by Kahneman and Tversky (1979) Won them the Nobel prize in 2002 Cognitive psychology basis for analyzing human errors
and deviations from rational behavior
Two important observations: Weighting effect: Players can subjectively weight
outcomes that are uncertain or risky Framing effect: Players may evaluate their utilities as
gains/losses with respect to a reference point
177
Illustrating the Weighting Effect
Weighting effect Prelec function
Outcomes are weighted
differently Weighting applies
toprobabilistic
outcomes (e.g. mixed strategies)
178
Prospect Theory
With weighting, the players now optimize:
Framing effect Each player will “frame” its gains/losses with
respect to a reference point Losses loom larger than gains
Weighting effect, Prelec function:
179
Prospect Theory
Concave in gains
Convex in losses
Steeper slope for losses as
opposed to gains Risk averse in
gains, risk seeking in losses
180
Prospect Theory
Framing effects The following framing function has been proposed:
Suitable applications for PT? When humans are making decisions (CPS with human-in-
the loop, smart grid, pricing , human hackers, security) UAV security is a prime example, given the impact of
UAV performance on owners/humans
Prospect Theory in UAV
The standard formulation does not account for the presence of humans in the loop that are facing uncertainty
Uncertainty: perceptions of both attacker and vendor on the probability of successful attack (weighting effect)
Framing: subjective perception on the delivery time with respect to a reference point Even the smallest of delays can be catastrophic For rescue situations, survival is at stake For Amazon, reputation can be damaged
181
Prospect Theory
Subjective, PT-based utility
The game is no longer zero-sum We consider max-min/min-max strategies
Ongoing work to characterize equilibria under PT182
Reference pointFraming function Weighting
183
Simulation results (1)
Due to the weightingeffect the vendorwill still choose
the shortestpath despite being
very risky (pn = 0.8)This choice
becomes more likelyas the vendorbecomes more
irrational
184
Simulation results (2)
Due to the weightingeffect, the attacker
focuses moreon nodes 5 and 8
which arepart of the shortest
path
185
Simulation results (3)
Delivery time is increased by almost 10%not accounting for time to re-load and
re-ship
186
Simulation results (4)
As the loss parameter increases
the vendor exaggerateslosses and thus
starts choosing morerisky paths to
meet delivery timewhich, in turn,
yields to a reverse effect!!!
Acknowledgment
Acknowledgement to students and collaborators: Mohammad Mozaffari, Anibal Sanjab, Mehdi Bennis, MingzheChen, Ursula Challita, Ismail Guvenc, Mérouane Debbah, and others
Acknowledgement to funding agencies: NSF and ONR
187
Conclusions UAVs provide with many new opportunities to
improve wireless communications The Internet of Flying Things will be upcoming and
we must be “analytically” ready Fundamental results on performance are needed Self-organization in terms of resources, network
topology, access modes, security, etc. Machine learning, game theory and related techniques
Human-in-the-loop: bounded rationality Ubiquitous wireless connectivity from the sky!
188
189
Finally….Thank YouQuestions?
References M. Mozaffari, W. Saad, M. Bennis, Y.-H. Nam, and M. Debbah, "A Tutorial on
UAVs for Wireless Networks: Applications, Challenges, and Open Problems", arXiv:1803.00680, 2018. https://arxiv.org/pdf/1803.00680.pdf
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial vehicle with underlaid device-to-device communications: performance and tradeoffs,” IEEE Transactions on Wireless Communications, vol. 15, no. 6, pp. 3949-3963, June 2016. available online : https://arxiv.org/pdf/1509.01187.pdf
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deployment of multiple unmanned aerial vehicles for optimal wireless coverage,” IEEE Communications Letters, vol. 20, no. 8, pp. 1647-1650, Aug. 2016. available online : https://arxiv.org/pdf/1606.01962.pdf
M. Mozaffari, A. T. Z. Kasgari, W. Saad, M. Bennis, and M. Debbah, "Beyond 5G with UAVs: Foundations of a 3D Wireless Cellular Network", arXiv:1805.06532, 2018. https://arxiv.org/pdf/1805.06532.pdf
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Optimal Transport Theory for Cell Association in UAV-Enabled Cellular Networks,” IEEE Communication Letters, to appear, June 2017. available online : https://arxiv.org/pdf/1705.09748.pdf
190
References M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Mobile Unmanned Aerial
Vehicles (UAVs) for Energy-Efficient Internet of Things Communications,” IEEE Transactions on Wireless Communications, 2017, available online : https://arxiv.org/pdf/1703.05401.pdf
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Wireless Communication using Unmanned Aerial Vehicles (UAVs): Optimal Transport Theory for Hover Time Optimization,” IEEE Transactions on Wireless Communications, 2017, available online : https://arxiv.org/pdf/1704.04813.pdf
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Mobile Internet of Things: Can UAVs provide an energy-effcient mobile architecture?,” in Proc. of IEEE Global Communications Conference (GLOBECOM), Washington, DC, USA, Dec. 2016. available online : https://arxiv.org/pdf/1607.02766.pdf
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Optimal transport theory for power efficient deployment of unmanned aerial vehicles," in Proc. of IEEE International Conference on Communications (ICC), Kuala Lumpur, May 2016. available online : https://arxiv.org/pdf/1602.01532.pdf
191
References M. Chen, M. Mozaffari, W. Saad, C. Yin, M. Debbah, and C. S. Hong, “Caching in the
Sky: Proactive Deployment of Cache-Enabled Unmanned Aerial Vehicles for Optimized Quality-of-Experience,” IEEE Journal on Selected Areas in Communications (JSAC), Special Issue on Human-In-The-Loop Mobile Networks, vol. 35, no. 5, pp. 1046-1061, May 2017. available online: https://arxiv.org/pdf/1610.01585.pdf
M. Naderisoorki, M. Mozaffari, H. Manshaei, and H. Saidi, “Resource Allocation for Machine-to-Machine Communications with Unmanned Aerial Vehicles," in Proc. of IEEE GLOBECOM Workshop, Washington, DC, USA, Dec. 2016. available online : https://arxiv.org/pdf/1608.07632.pdf
D. Athukoralage, I. Guvenc, W. Saad, and M. Bennis, "Regret Based Learning for UAV assisted LTE-U/WiFi Public Safety Networks," in Proc. of the IEEE Global Communications Conference (GLOBECOM), Mobile and Wireless Networks Symposium, Washington, DC, USA, Dec. 2016. available online : http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7842208
A. Sanjab, W. Saad, and T. Başar, "Prospect Theory for Enhanced Cyber-Physical Security of Drone Delivery Systems: A Network Interdiction Game," in Proc. of the IEEE International Conference on Communications (ICC), Communication and Information Systems Security Symposium, Paris, France, May 2017. available online : https://arxiv.org/pdf/1702.04240.pdf 192
References W. Saad, Z. Han, T. Başar, M. Debbah, and A. Hjorungnes, “Hedonic Coalition
Formation for Distributed Task Allocation among Wireless Agents,” IEEE Transactions on Mobile Computing, vol. 10, no.9, pp. 1327-1344, Sep. 2011. available online : https://arxiv.org/pdf/1010.4499.pdf
U. Challita and W. Saad, "Network Formation in the Sky: Unmanned Aerial Vehicles for Multi-hop Wireless Backhauling", in Proc. of the IEEE Global Communications Conference (GLOBECOM), Singapore, December 2017.
M. Chen, W. Saad, and C. Yin, "Liquid State Machine Learning for Resource Allocation in a Network of Cache-Enabled LTE-U UAVs", in Proc. of the IEEE Global Communications Conference (GLOBECOM), Singapore, December 2017.
A. Hourani, S. Kandeepan, and A. Jamalipour, “Modeling air-to-ground path loss for low altitude platforms in urban environments,” in Proc. of IEEE Global Telecommunications Conference, Austin, Tx, USA, Dec. 2014
M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Drone small cells in the clouds: Design, deployment and performance analysis,” in Proc. of IEEE Global Communications Conference (GLOBECOM), San Diego, CA, USA, Dec. 2015. available online : https://arxiv.org/pdf/1509.01655.pdf
193